{"id":12658,"date":"2023-02-04T17:42:00","date_gmt":"2023-02-04T09:42:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=12658"},"modified":"2023-02-03T17:42:18","modified_gmt":"2023-02-03T09:42:18","slug":"torchscript","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/02\/04\/torchscript\/","title":{"rendered":"TorchScript&#8212;\u6a21\u578b\u90e8\u7f72"},"content":{"rendered":"\n<p>\u6458\u81ea\uff1a<a href=\"https:\/\/zhuanlan.zhihu.com\/p\/486914187\">https:\/\/zhuanlan.zhihu.com\/p\/486914187<\/a><\/p>\n\n\n\n<p class=\"has-text-align-center has-light-pink-background-color has-background\">\u5b98\u7f51\uff1ahttps:\/\/pytorch.org\/docs\/stable\/jit.html<\/p>\n\n\n\n<p>PyTorch \u65e0\u7591\u662f\u73b0\u5728\u6700\u6210\u529f\u7684\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u6846\u67b6\u4e4b\u4e00\uff0c\u662f\u5404\u79cd\u9876\u4f1a\u9876\u520a\u8bba\u6587\u5b9e\u9a8c\u7684\u5927\u70ed\u95e8\u3002\u6bd4\u8d77\u5176\u4ed6\u7684\u6846\u67b6\uff0cPyTorch \u6700\u5927\u7684\u5356\u70b9\u662f\u5b83\u5bf9\u52a8\u6001\u7f51\u7edc\u7684\u652f\u6301\uff0c\u6bd4\u5176\u4ed6\u9700\u8981\u6784\u5efa\u9759\u6001\u7f51\u7edc\u7684\u6846\u67b6\u62e5\u6709\u66f4\u4f4e\u7684\u5b66\u4e60\u6210\u672c\u3002PyTorch \u6e90\u7801 Readme \u4e2d\u8fd8\u4e13\u95e8\u4e3a\u6b64\u505a\u4e86\u4e00\u5f20\u52a8\u6001\u56fe\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"262\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-21-1024x262.png\" alt=\"\" class=\"wp-image-12662\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-21-1024x262.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-21-300x77.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-21-768x197.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/02\/image-21.png 1227w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><br>\u5bf9\u7814\u7a76\u5458\u800c\u8a00\uff0c PyTorch \u80fd\u6781\u5927\u5730\u63d0\u9ad8\u60f3 idea\u3001\u505a\u5b9e\u9a8c\u3001\u53d1\u8bba\u6587\u7684\u6548\u7387\uff0c\u662f\u8bad\u7ec3\u6846\u67b6\u4e2d\u7684\u8c6a\u6770\uff0c\u4f46\u662f\u5b83\u4e0d\u9002\u5408\u90e8\u7f72\u3002\u52a8\u6001\u5efa\u56fe\u5e26\u6765\u7684\u4f18\u52bf\u5bf9\u4e8e\u6027\u80fd\u8981\u6c42\u66f4\u9ad8\u7684\u5e94\u7528\u573a\u666f\u800c\u8a00\u66f4\u50cf\u662f\u7f3a\u70b9\uff0c\u975e\u56fa\u5b9a\u7684\u7f51\u7edc\u7ed3\u6784\u7ed9\u7f51\u7edc\u7ed3\u6784\u5206\u6790\u5e76\u8fdb\u884c\u4f18\u5316\u5e26\u6765\u4e86\u56f0\u96be\uff0c\u591a\u6570\u53c2\u6570\u90fd\u80fd\u4ee5 Tensor \u5f62\u5f0f\u4f20\u8f93\u4e5f\u8ba9\u8d44\u6e90\u5206\u914d\u53d8\u6210\u4e00\u4ef6\u95f9\u5fc3\u7684\u4e8b\u3002\u53e6\u5916\u7531\u4e8e\u56fe\u662f\u7531 python \u4ee3\u7801\u6784\u5efa\u7684\uff0c\u4e00\u65b9\u9762\u90e8\u7f72\u8981\u4f9d\u8d56 python \u73af\u5883\uff0c\u53e6\u4e00\u65b9\u9762\u6a21\u578b\u4e5f\u6beb\u65e0\u4fdd\u5bc6\u6027\u53ef\u8a00\u3002<\/p>\n\n\n\n<p><strong>\u800c TorchScript \u5c31\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u800c\u8bde\u751f\u7684\u5de5\u5177\u3002\u5305\u62ec\u4ee3\u7801\u7684\u8ffd\u8e2a\u53ca\u89e3\u6790\u3001\u4e2d\u95f4\u8868\u793a\u7684\u751f\u6210\u3001\u6a21\u578b\u4f18\u5316\u3001\u5e8f\u5217\u5316\u7b49\u5404\u79cd\u529f\u80fd\uff0c\u53ef\u4ee5\u8bf4\u662f\u8986\u76d6\u4e86\u6a21\u578b\u90e8\u7f72\u7684\u65b9\u65b9\u9762\u9762\u3002<\/strong><\/p>\n\n\n\n<h2><strong>TorchScript<\/strong><\/h2>\n\n\n\n<p>\u52a8\u6001\u56fe\u6a21\u578b\u901a\u8fc7\u727a\u7272\u4e00\u4e9b\u9ad8\u7ea7\u7279\u6027\u6765\u6362\u53d6\u6613\u7528\u6027\uff0c\u90a3\u5230\u5e95 JIT \u6709\u54ea\u4e9b\u7279\u6027\uff0c\u5728\u4ec0\u4e48\u60c5\u51b5\u4e0b\u4e0d\u5f97\u4e0d\u7528\u5230 JIT \u5462\uff1f\u4e0b\u9762\u4e3b\u8981\u901a\u8fc7\u4ecb\u7ecd TorchScript\uff08PyTorch \u7684 JIT \u5b9e\u73b0\uff09\u6765\u5206\u6790 JIT \u5230\u5e95\u5e26\u6765\u4e86\u54ea\u4e9b\u597d\u5904\u3002<\/p>\n\n\n\n<ol><li>\u6a21\u578b\u90e8\u7f72<\/li><\/ol>\n\n\n\n<p>PyTorch \u7684 1.0 \u7248\u672c\u53d1\u5e03\u7684\u6700\u6838\u5fc3\u7684\u4e24\u4e2a\u65b0\u7279\u6027\u5c31\u662f JIT \u548c C++ API\uff0c\u8fd9\u4e24\u4e2a\u7279\u6027\u4e00\u8d77\u53d1\u5e03\u4e0d\u662f\u6ca1\u6709\u9053\u7406\u7684\uff0cJIT \u662f Python \u548c C++ \u7684\u6865\u6881\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528 Python \u8bad\u7ec3\u6a21\u578b\uff0c\u7136\u540e\u901a\u8fc7 JIT \u5c06\u6a21\u578b\u8f6c\u4e3a\u8bed\u8a00\u65e0\u5173\u7684\u6a21\u5757\uff0c\u4ece\u800c\u8ba9 C++ \u53ef\u4ee5\u975e\u5e38\u65b9\u4fbf\u5f97\u8c03\u7528\uff0c\u4ece\u6b64\u300c\u4f7f\u7528 Python \u8bad\u7ec3\u6a21\u578b\uff0c\u4f7f\u7528 C++ \u5c06\u6a21\u578b\u90e8\u7f72\u5230\u751f\u4ea7\u73af\u5883\u300d\u5bf9 PyTorch \u6765\u8bf4\u6210\u4e3a\u4e86\u4e00\u4ef6\u5f88\u5bb9\u6613\u7684\u4e8b\u3002\u800c\u56e0\u4e3a\u4f7f\u7528\u4e86 C++\uff0c\u6211\u4eec\u73b0\u5728\u51e0\u4e4e\u53ef\u4ee5\u628a PyTorch \u6a21\u578b\u90e8\u7f72\u5230\u4efb\u610f\u5e73\u53f0\u548c\u8bbe\u5907\u4e0a\uff1a\u6811\u8393\u6d3e\u3001iOS\u3001Android \u7b49\u7b49\u2026<\/p>\n\n\n\n<p>2. \u6027\u80fd\u63d0\u5347<\/p>\n\n\n\n<p>\u65e2\u7136\u662f\u4e3a\u90e8\u7f72\u751f\u4ea7\u6240\u63d0\u4f9b\u7684\u7279\u6027\uff0c\u90a3\u514d\u4e0d\u4e86\u5728\u6027\u80fd\u4e0a\u9762\u505a\u4e86\u6781\u5927\u7684\u4f18\u5316\uff0c\u5982\u679c\u63a8\u65ad\u7684\u573a\u666f\u5bf9\u6027\u80fd\u8981\u6c42\u9ad8\uff0c\u5219\u53ef\u4ee5\u8003\u8651\u5c06\u6a21\u578b\uff08torch.nn.Module\uff09\u8f6c\u6362\u4e3a TorchScript Module\uff0c\u518d\u8fdb\u884c\u63a8\u65ad\u3002<\/p>\n\n\n\n<p>3. \u6a21\u578b\u53ef\u89c6\u5316<\/p>\n\n\n\n<p>TensorFlow \u6216 Keras \u5bf9\u6a21\u578b\u53ef\u89c6\u5316\u5de5\u5177\uff08TensorBoard\u7b49\uff09\u975e\u5e38\u53cb\u597d\uff0c\u56e0\u4e3a\u672c\u8eab\u5c31\u662f\u9759\u6001\u56fe\u7684\u7f16\u7a0b\u6a21\u578b\uff0c\u5728\u6a21\u578b\u5b9a\u4e49\u597d\u540e\u6574\u4e2a\u6a21\u578b\u7684\u7ed3\u6784\u548c\u6b63\u5411\u903b\u8f91\u5c31\u5df2\u7ecf\u6e05\u695a\u4e86\uff1b\u4f46 PyTorch \u672c\u8eab\u662f\u4e0d\u652f\u6301\u7684\uff0c\u6240\u4ee5 PyTorch \u6a21\u578b\u5728\u53ef\u89c6\u5316\u4e0a\u4e00\u76f4\u8868\u73b0\u5f97\u4e0d\u597d\uff0c\u4f46 JIT \u6539\u5584\u4e86\u8fd9\u4e00\u60c5\u51b5\u3002\u73b0\u5728\u53ef\u4ee5\u4f7f\u7528 JIT \u7684 trace \u529f\u80fd\u6765\u5f97\u5230 PyTorch \u6a21\u578b\u9488\u5bf9\u67d0\u4e00\u8f93\u5165\u7684\u6b63\u5411\u903b\u8f91\uff0c\u901a\u8fc7\u6b63\u5411\u903b\u8f91\u53ef\u4ee5\u5f97\u5230\u6a21\u578b\u5927\u81f4\u7684\u7ed3\u6784\uff0c\u4f46\u5982\u679c\u5728 `forward` \u65b9\u6cd5\u4e2d\u6709\u5f88\u591a\u6761\u4ef6\u63a7\u5236\u8bed\u53e5\uff0c\u8fd9\u4f9d\u7136\u4e0d\u662f\u4e00\u4e2a\u597d\u7684\u65b9\u6cd5\uff0c\u6240\u4ee5 PyTorch JIT \u8fd8\u63d0\u4f9b\u4e86 Scripting \u7684\u65b9\u5f0f\u3002<\/p>\n\n\n\n<h2><strong>TorchScript Module&nbsp;<\/strong>\u7684\u4e24\u79cd\u751f\u6210\u65b9\u5f0f<\/h2>\n\n\n\n<p><strong>1. \u7f16\u7801\uff08Scripting\uff09<\/strong><\/p>\n\n\n\n<p>\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528 TorchScript Language \u6765\u5b9a\u4e49\u4e00\u4e2a PyTorch JIT Module\uff0c\u7136\u540e\u7528&nbsp;<strong>torch.jit.script<\/strong>&nbsp;\u6765\u5c06\u4ed6\u8f6c\u6362\u6210 TorchScript Module \u5e76\u4fdd\u5b58\u6210\u6587\u4ef6\u3002\u800c TorchScript Language \u672c\u8eab\u4e5f\u662f Python \u4ee3\u7801\uff0c\u6240\u4ee5\u53ef\u4ee5\u76f4\u63a5\u5199\u5728 Python \u6587\u4ef6\u4e2d\u3002<\/p>\n\n\n\n<p>\u4f7f\u7528 TorchScript Language \u5c31\u5982\u540c\u4f7f\u7528 TensorFlow \u4e00\u6837\uff0c\u9700\u8981\u524d\u5b9a\u4e49\u597d\u5b8c\u6574\u7684\u56fe\u3002\u5bf9\u4e8e TensorFlow \u6211\u4eec\u77e5\u9053\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528 Python \u4e2d\u7684 if \u7b49\u8bed\u53e5\u6765\u505a\u6761\u4ef6\u63a7\u5236\uff0c\u800c\u662f\u9700\u8981\u7528 tf.cond\uff0c\u4f46\u5bf9\u4e8e TorchScript \u6211\u4eec\u4f9d\u7136\u80fd\u591f\u76f4\u63a5\u4f7f\u7528 if \u548c for \u7b49\u6761\u4ef6\u63a7\u5236\u8bed\u53e5\uff0c\u6240\u4ee5\u5373\u4f7f\u662f\u5728\u9759\u6001\u56fe\u4e0a\uff0cPyTorch \u4f9d\u7136\u79c9\u627f\u4e86\u300c\u6613\u7528\u300d\u7684\u7279\u6027\u3002TorchScript Language \u662f\u9759\u6001\u7c7b\u578b\u7684 Python \u5b50\u96c6\uff0c\u9759\u6001\u7c7b\u578b\u4e5f\u662f\u7528\u4e86 Python 3 \u7684 typing \u6a21\u5757\u6765\u5b9e\u73b0\uff0c\u6240\u4ee5\u5199 TorchScript Language \u7684\u4f53\u9a8c\u4e5f\u8ddf Python \u4e00\u6a21\u4e00\u6837\uff0c\u53ea\u662f\u67d0\u4e9b Python \u7279\u6027\u65e0\u6cd5\u4f7f\u7528\uff08\u56e0\u4e3a\u662f\u5b50\u96c6\uff09\uff0c\u53ef\u4ee5\u901a\u8fc7&nbsp;<a href=\"https:\/\/pytorch.org\/docs\/stable\/jit_language_reference.html\" target=\"_blank\" rel=\"noreferrer noopener\">TorchScript Language Reference<\/a>&nbsp;\u6765\u67e5\u770b\u548c\u539f\u751f Python \u7684\u5f02\u540c\u3002<\/p>\n\n\n\n<p>\u7406\u8bba\u4e0a\uff0c\u4f7f\u7528 Scripting \u7684\u65b9\u5f0f\u5b9a\u4e49\u7684 TorchScript Module \u5bf9\u6a21\u578b\u53ef\u89c6\u5316\u5de5\u5177\u975e\u5e38\u53cb\u597d\uff0c\u56e0\u4e3a\u5df2\u7ecf\u63d0\u524d\u5b9a\u4e49\u4e86\u6574\u4e2a\u56fe\u7ed3\u6784\u3002<\/p>\n\n\n\n<p><strong>2. \u8ffd\u8e2a\uff08Tracing\uff09<\/strong><\/p>\n\n\n\n<p>\u4f7f\u7528 TorchScript Module \u7684\u66f4\u7b80\u5355\u7684\u529e\u6cd5\u662f\u4f7f\u7528 Tracing\uff0cTracing \u53ef\u4ee5\u76f4\u63a5\u5c06 PyTorch \u6a21\u578b\uff08torch.nn.Module\uff09\u8f6c\u6362\u6210 TorchScript Module\u3002\u300c\u8ffd\u8e2a\u300d\u987e\u540d\u601d\u4e49\uff0c\u5c31\u662f\u9700\u8981\u63d0\u4f9b\u4e00\u4e2a\u300c\u8f93\u5165\u300d\u6765\u8ba9\u6a21\u578b forward \u4e00\u904d\uff0c\u4ee5\u901a\u8fc7\u8be5\u8f93\u5165\u7684\u6d41\u8f6c\u8def\u5f84\uff0c\u83b7\u5f97\u56fe\u7684\u7ed3\u6784\u3002\u8fd9\u79cd\u65b9\u5f0f\u5bf9\u4e8e forward \u903b\u8f91\u7b80\u5355\u7684\u6a21\u578b\u6765\u8bf4\u975e\u5e38\u5b9e\u7528\uff0c\u4f46\u5982\u679c forward \u91cc\u9762\u672c\u8eab\u5939\u6742\u4e86\u5f88\u591a\u6d41\u7a0b\u63a7\u5236\u8bed\u53e5\uff0c\u5219\u53ef\u80fd\u4f1a\u6709\u95ee\u9898\uff0c\u56e0\u4e3a\u540c\u4e00\u4e2a\u8f93\u5165\u4e0d\u53ef\u80fd\u904d\u5386\u5230\u6240\u6709\u7684\u903b\u8f91\u5206\u679d\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u8fd8\u53ef\u4ee5\u6df7\u5408\u4f7f\u7528\u4e0a\u9762\u4e24\u79cd\u65b9\u5f0f\u3002<\/p>\n\n\n\n<h2 id=\"h_486914187_1\">\u6a21\u578b\u8f6c\u6362<\/h2>\n\n\n\n<p>\u4f5c\u4e3a\u6a21\u578b\u90e8\u7f72\u7684\u4e00\u4e2a\u8303\u5f0f\uff0c\u901a\u5e38\u6211\u4eec\u90fd\u9700\u8981\u751f\u6210\u4e00\u4e2a\u6a21\u578b\u7684\u4e2d\u95f4\u8868\u793a\uff08IR\uff09\uff0c\u8fd9\u4e2a IR \u62e5\u6709\u76f8\u5bf9\u56fa\u5b9a\u7684\u56fe\u7ed3\u6784\uff0c\u6240\u4ee5\u66f4\u5bb9\u6613\u4f18\u5316\uff0c\u8ba9\u6211\u4eec\u770b\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch \nfrom torchvision.models import resnet18 \n \n<em># \u4f7f\u7528PyTorch model zoo\u4e2d\u7684resnet18\u4f5c\u4e3a\u4f8b\u5b50 <\/em>\nmodel = resnet18() \nmodel.eval() \n \n<em># \u901a\u8fc7trace\u7684\u65b9\u6cd5\u751f\u6210IR\u9700\u8981\u4e00\u4e2a\u8f93\u5165\u6837\u4f8b <\/em>\ndummy_input = torch.rand(1, 3, 224, 224) \n \n<em># IR\u751f\u6210 <\/em>\nwith torch.no_grad(): \n    jit_model = torch.jit.trace(model, dummy_input) <\/code><\/pre>\n\n\n\n<p>JIT \u662f\u4e00\u79cd\u6982\u5ff5\uff0c\u5168\u79f0\u662f Just In Time Compilation\uff0c\u4e2d\u6587\u8bd1\u4e3a\u300c\u5373\u65f6\u7f16\u8bd1\u300d\uff0c\u662f\u4e00\u79cd\u7a0b\u5e8f\u4f18\u5316\u7684\u65b9\u6cd5<\/p>\n\n\n\n<p>\u5230\u8fd9\u91cc\u5c31\u5c06 PyTorch \u7684\u6a21\u578b\u8f6c\u6362\u6210\u4e86 TorchScript \u7684 IR\u3002\u8fd9\u91cc\u6211\u4eec\u4f7f\u7528\u4e86 trace \u6a21\u5f0f\u6765\u751f\u6210 IR\uff0c\u6240\u8c13 trace \u6307\u7684\u662f\u8fdb\u884c\u4e00\u6b21\u6a21\u578b\u63a8\u7406\uff0c\u5728\u63a8\u7406\u7684\u8fc7\u7a0b\u4e2d\u8bb0\u5f55\u6240\u6709\u7ecf\u8fc7\u7684\u8ba1\u7b97\uff0c\u5c06\u8fd9\u4e9b\u8bb0\u5f55\u6574\u5408\u6210\u8ba1\u7b97\u56fe\u3002<\/p>\n\n\n\n<p>\u90a3\u4e48\u8fd9\u4e2a IR \u4e2d\u5230\u5e95\u90fd\u6709\u4e9b\u4ec0\u4e48\u5462\uff1f\u6211\u4eec\u53ef\u4ee5\u53ef\u89c6\u5316\u4e00\u4e0b\u5176\u4e2d\u7684 layer1 \u770b\u770b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>jit_layer1 = jit_model.layer1 \nprint(jit_layer1.graph) \n \n<em># graph(%self.6 : __torch__.torch.nn.modules.container.Sequential, <\/em>\n<em>#       %4 : Float(1, 64, 56, 56, strides=&#91;200704, 3136, 56, 1], requires_grad=0, device=cpu)): <\/em>\n<em>#   %1 : __torch__.torchvision.models.resnet.___torch_mangle_10.BasicBlock = prim::GetAttr&#91;name=\"1\"](%self.6) <\/em>\n<em>#   %2 : __torch__.torchvision.models.resnet.BasicBlock = prim::GetAttr&#91;name=\"0\"](%self.6) <\/em>\n<em>#   %6 : Tensor = prim::CallMethod&#91;name=\"forward\"](%2, %4) <\/em>\n<em>#   %7 : Tensor = prim::CallMethod&#91;name=\"forward\"](%1, %6) <\/em>\n<em>#   return (%7) <\/em><\/code><\/pre>\n\n\n\n<p>\u662f\u4e0d\u662f\u6709\u70b9\u6478\u4e0d\u7740\u5934\u8111\uff1fTorchScript \u6709\u5b83\u81ea\u5df1\u5bf9\u4e8e Graph \u4ee5\u53ca\u5176\u4e2d\u5143\u7d20\u7684\u5b9a\u4e49\uff0c\u5bf9\u4e8e\u7b2c\u4e00\u6b21\u63a5\u89e6\u7684\u4eba\u6765\u8bf4\u53ef\u80fd\u6bd4\u8f83\u964c\u751f\uff0c\u4f46\u662f\u6ca1\u5173\u7cfb\uff0c\u6211\u4eec\u8fd8\u6709\u53e6\u4e00\u79cd\u53ef\u89c6\u5316\u65b9\u5f0f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>print(jit_layer1.code) \n \n<em># def forward(self, <\/em>\n<em>#     argument_1: Tensor) -&gt; Tensor: <\/em>\n<em>#   _0 = getattr(self, \"1\") <\/em>\n<em>#   _1 = (getattr(self, \"0\")).forward(argument_1, ) <\/em>\n<em>#   return (_0).forward(_1, ) <\/em><\/code><\/pre>\n\n\n\n<p>\u6ca1\u9519\uff0c\u5c31\u662f\u4ee3\u7801\uff01TorchScript \u7684 IR \u662f\u53ef\u4ee5\u8fd8\u539f\u6210 python \u4ee3\u7801\u7684\uff0c\u5982\u679c\u4f60\u751f\u6210\u4e86\u4e00\u4e2a TorchScript \u6a21\u578b\u5e76\u4e14\u60f3\u77e5\u9053\u5b83\u7684\u5185\u5bb9\u5bf9\u4e0d\u5bf9\uff0c\u90a3\u4e48\u53ef\u4ee5\u901a\u8fc7\u8fd9\u6837\u7684\u65b9\u5f0f\u6765\u505a\u4e00\u4e9b\u7b80\u5355\u7684\u68c0\u67e5\u3002<\/p>\n\n\n\n<p>\u521a\u624d\u7684\u4f8b\u5b50\u4e2d\u6211\u4eec\u4f7f\u7528 trace \u7684\u65b9\u6cd5\u751f\u6210IR\u3002\u9664\u4e86 trace \u4e4b\u5916\uff0cPyTorch \u8fd8\u63d0\u4f9b\u4e86\u53e6\u4e00\u79cd\u751f\u6210 TorchScript \u6a21\u578b\u7684\u65b9\u6cd5\uff1ascript\u3002\u8fd9\u79cd\u65b9\u5f0f\u4f1a\u76f4\u63a5\u89e3\u6790\u7f51\u7edc\u5b9a\u4e49\u7684 python \u4ee3\u7801\uff0c\u751f\u6210\u62bd\u8c61\u8bed\u6cd5\u6811 AST\uff0c\u56e0\u6b64\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u89e3\u51b3\u4e00\u4e9b trace \u65e0\u6cd5\u89e3\u51b3\u7684\u95ee\u9898\uff0c\u6bd4\u5982\u5bf9 branch\/loop \u7b49\u6570\u636e\u6d41\u63a7\u5236\u8bed\u53e5\u7684\u5efa\u56fe\u3002script\u65b9\u5f0f\u7684\u5efa\u56fe\u6709\u5f88\u591a\u6709\u8da3\u7684\u7279\u6027\uff0c\u4f1a\u5728\u672a\u6765\u7684\u5206\u4eab\u4e2d\u505a\u4e13\u9898\u5206\u6790\uff0c\u656c\u8bf7\u671f\u5f85\u3002<\/p>\n\n\n\n<h2 id=\"h_486914187_2\">\u6a21\u578b\u4f18\u5316<\/h2>\n\n\n\n<p>\u806a\u660e\u7684\u540c\u5b66\u53ef\u80fd\u53d1\u73b0\u4e86\uff0c\u4e0a\u9762\u7684\u53ef\u89c6\u5316\u4e2d\u53ea\u6709<code>resnet18<\/code>\u91cc<code>forward<\/code>\u7684\u90e8\u5206\uff0c\u5176\u4e2d\u7684\u5b50\u6a21\u5757\u4fe1\u606f\u662f\u4e0d\u662f\u4e22\u5931\u4e86\u5462\uff1f\u5982\u679c\u6ca1\u6709\u4e22\u5931\uff0c\u90a3\u4e48\u600e\u4e48\u6837\u624d\u80fd\u786e\u5b9a\u5b50\u6a21\u5757\u7684\u5185\u5bb9\u662f\u5426\u6b63\u786e\u5462\uff1f\u522b\u62c5\u5fc3\uff0c\u8fd8\u8bb0\u5f97\u6211\u4eec\u8bf4\u8fc7 TorchScript \u652f\u6301\u5bf9\u7f51\u7edc\u7684\u4f18\u5316\u5417\uff0c\u8fd9\u91cc\u6211\u4eec\u5c31\u53ef\u4ee5\u7528\u4e00\u4e2a<code>pass<\/code>\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code><em># \u8c03\u7528inline pass\uff0c\u5bf9graph\u505a\u53d8\u6362 <\/em>\ntorch._C._jit_pass_inline(jit_layer1.graph) \nprint(jit_layer1.code) \n \n<em># def forward(self, <\/em>\n<em>#     argument_1: Tensor) -&gt; Tensor: <\/em>\n<em>#   _0 = getattr(self, \"1\") <\/em>\n<em>#   _1 = getattr(self, \"0\") <\/em>\n<em>#   _2 = _1.bn2 <\/em>\n<em>#   _3 = _1.conv2 <\/em>\n<em>#   _4 = _1.bn1 <\/em>\n<em>#   input = torch._convolution(argument_1, _1.conv1.weight, None, &#91;1, 1], &#91;1, 1], &#91;1, 1], False, &#91;0, 0], 1, False, False, True, True) <\/em>\n<em>#   _5 = _4.running_var <\/em>\n<em>#   _6 = _4.running_mean <\/em>\n<em>#   _7 = _4.bias <\/em>\n<em>#   input0 = torch.batch_norm(input, _4.weight, _7, _6, _5, False, 0.10000000000000001, 1.0000000000000001e-05, True) <\/em>\n<em>#   input1 = torch.relu_(input0) <\/em>\n<em>#   input2 = torch._convolution(input1, _3.weight, None, &#91;1, 1], &#91;1, 1], &#91;1, 1], False, &#91;0, 0], 1, False, False, True, True) <\/em>\n<em>#   _8 = _2.running_var <\/em>\n<em>#   _9 = _2.running_mean <\/em>\n<em>#   _10 = _2.bias <\/em>\n<em>#   out = torch.batch_norm(input2, _2.weight, _10, _9, _8, False, 0.10000000000000001, 1.0000000000000001e-05, True) <\/em>\n<em>#   input3 = torch.add_(out, argument_1, alpha=1) <\/em>\n<em>#   input4 = torch.relu_(input3) <\/em>\n<em>#   _11 = _0.bn2 <\/em>\n<em>#   _12 = _0.conv2 <\/em>\n<em>#   _13 = _0.bn1 <\/em>\n<em>#   input5 = torch._convolution(input4, _0.conv1.weight, None, &#91;1, 1], &#91;1, 1], &#91;1, 1], False, &#91;0, 0], 1, False, False, True, True) <\/em>\n<em>#   _14 = _13.running_var <\/em>\n<em>#   _15 = _13.running_mean <\/em>\n<em>#   _16 = _13.bias <\/em>\n<em>#   input6 = torch.batch_norm(input5, _13.weight, _16, _15, _14, False, 0.10000000000000001, 1.0000000000000001e-05, True) <\/em>\n<em>#   input7 = torch.relu_(input6) <\/em>\n<em>#   input8 = torch._convolution(input7, _12.weight, None, &#91;1, 1], &#91;1, 1], &#91;1, 1], False, &#91;0, 0], 1, False, False, True, True) <\/em>\n<em>#   _17 = _11.running_var <\/em>\n<em>#   _18 = _11.running_mean <\/em>\n<em>#   _19 = _11.bias <\/em>\n<em>#   out0 = torch.batch_norm(input8, _11.weight, _19, _18, _17, False, 0.10000000000000001, 1.0000000000000001e-05, True) <\/em>\n<em>#   input9 = torch.add_(out0, input4, alpha=1) <\/em>\n<em>#   return torch.relu_(input9) <\/em><\/code><\/pre>\n\n\n\n<p>\u8fd9\u91cc\u6211\u4eec\u5c31\u80fd\u770b\u5230\u5377\u79ef\u3001batch_norm\u3001relu\u7b49\u719f\u6089\u7684\u7b97\u5b50\u4e86\u3002<\/p>\n\n\n\n<p>\u4e0a\u9762\u4ee3\u7801\u4e2d\u6211\u4eec\u4f7f\u7528\u4e86\u4e00\u4e2a\u540d\u4e3a<code>inline<\/code>\u7684<code>pass<\/code>\uff0c\u5c06\u6240\u6709\u5b50\u6a21\u5757\u8fdb\u884c\u5185\u8054\uff0c\u8fd9\u6837\u6211\u4eec\u5c31\u80fd\u770b\u89c1\u66f4\u5b8c\u6574\u7684\u63a8\u7406\u4ee3\u7801\u3002<code>pass<\/code>\u662f\u4e00\u4e2a\u6765\u6e90\u4e8e\u7f16\u8bd1\u539f\u7406\u7684\u6982\u5ff5\uff0c\u4e00\u4e2a TorchScript \u7684 pass \u4f1a\u63a5\u6536\u4e00\u4e2a\u56fe\uff0c\u904d\u5386\u56fe\u4e2d\u6240\u6709\u5143\u7d20\u8fdb\u884c\u67d0\u79cd\u53d8\u6362\uff0c\u751f\u6210\u4e00\u4e2a\u65b0\u7684\u56fe\u3002\u6211\u4eec\u8fd9\u91cc\u7528\u5230\u7684<code>inline<\/code>\u8d77\u5230\u7684\u4f5c\u7528\u5c31\u662f\u5c06\u6a21\u5757\u8c03\u7528\u5c55\u5f00\uff0c\u5c3d\u7ba1\u8fd9\u6837\u505a\u5e76\u4e0d\u80fd\u76f4\u63a5\u5f71\u54cd\u6267\u884c\u6548\u7387\uff0c\u4f46\u662f\u5b83\u5176\u5b9e\u662f\u5f88\u591a\u5176\u4ed6<code>pass<\/code>\u7684\u57fa\u7840\u3002PyTorch \u4e2d\u5b9a\u4e49\u4e86\u975e\u5e38\u591a\u7684 pass \u6765\u89e3\u51b3\u5404\u79cd\u4f18\u5316\u4efb\u52a1\uff0c\u672a\u6765\u6211\u4eec\u4f1a\u505a\u4e00\u4e9b\u66f4\u8be6\u7ec6\u7684\u4ecb\u7ecd\u3002<\/p>\n\n\n\n<h2 id=\"h_486914187_3\">\u5e8f\u5217\u5316<\/h2>\n\n\n\n<p>\u4e0d\u7ba1\u662f\u54ea\u79cd\u65b9\u6cd5\u521b\u5efa\u7684 TorchScript \u90fd\u53ef\u4ee5\u8fdb\u884c\u5e8f\u5217\u5316\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code><em># \u5c06\u6a21\u578b\u5e8f\u5217\u5316 <\/em>\njit_model.save('jit_model.pth') \n<em># \u52a0\u8f7d\u5e8f\u5217\u5316\u540e\u7684\u6a21\u578b <\/em>\njit_model = torch.jit.load('jit_model.pth') <\/code><\/pre>\n\n\n\n<p>\u5e8f\u5217\u5316\u540e\u7684\u6a21\u578b\u4e0d\u518d\u4e0e python \u76f8\u5173\uff0c\u53ef\u4ee5\u88ab\u90e8\u7f72\u5230\u5404\u79cd\u5e73\u53f0\u4e0a\u3002<\/p>\n\n\n\n<p>PyTorch \u63d0\u4f9b\u4e86\u53ef\u4ee5\u7528\u4e8e TorchScript \u6a21\u578b\u63a8\u7406\u7684 c++ API\uff0c\u5e8f\u5217\u5316\u540e\u7684\u6a21\u578b\u7ec8\u4e8e\u53ef\u4ee5\u4e0d\u4f9d\u8d56 python \u8fdb\u884c\u63a8\u7406\u4e86\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code><em>\/\/ \u52a0\u8f7d\u751f\u6210\u7684torchscript\u6a21\u578b \n<\/em>auto module = torch::jit::load('jit_model.pth'); \n<em>\/\/ \u6839\u636e\u4efb\u52a1\u9700\u6c42\u8bfb\u53d6\u6570\u636e \n<\/em>std::vector&lt;torch::jit::IValue> inputs = ...; \n<em>\/\/ \u8ba1\u7b97\u63a8\u7406\u7ed3\u679c \n<\/em>auto output = module.forward(inputs).toTensor(); <\/code><\/pre>\n\n\n\n<p>\u4e0e torch.onnx \u7684\u5173\u7cfb\uff1aONNX \u662f\u4e1a\u754c\u5e7f\u6cdb\u4f7f\u7528\u7684\u4e00\u79cd\u795e\u7ecf\u7f51\u7edc\u4e2d\u95f4\u8868\u793a\uff0cPyTorch \u81ea\u7136\u4e5f\u5bf9 ONNX \u63d0\u4f9b\u4e86\u652f\u6301\u3002<code>torch.onnx.export<\/code>\u51fd\u6570\u53ef\u4ee5\u5e2e\u52a9\u6211\u4eec\u628a PyTorch \u6a21\u578b\u8f6c\u6362\u6210 ONNX \u6a21\u578b\uff0c\u8fd9\u4e2a\u51fd\u6570\u4f1a\u4f7f\u7528 trace \u7684\u65b9\u5f0f\u8bb0\u5f55 PyTorch \u7684\u63a8\u7406\u8fc7\u7a0b\u3002\u806a\u660e\u7684\u540c\u5b66\u53ef\u80fd\u5df2\u7ecf\u60f3\u5230\u4e86\uff0c\u6ca1\u9519\uff0cONNX \u7684\u5bfc\u51fa\uff0c\u4f7f\u7528\u7684\u6b63\u662f TorchScript \u7684 trace \u5de5\u5177\u3002\u5177\u4f53\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ol><li>\u4f7f\u7528 trace \u7684\u65b9\u5f0f\u5148\u751f\u6210\u4e00\u4e2a TorchScipt \u6a21\u578b\uff0c\u5982\u679c\u4f60\u8f6c\u6362\u7684\u672c\u8eab\u5c31\u662f TorchScript \u6a21\u578b\uff0c\u5219\u53ef\u4ee5\u8df3\u8fc7\u8fd9\u4e00\u6b65\u3002<\/li><li>\u4f7f\u7528\u8bb8\u591a pass \u5bf9 1 \u4e2d\u751f\u6210\u7684\u6a21\u578b\u8fdb\u884c\u53d8\u6362\uff0c\u5176\u4e2d\u5bf9 ONNX \u5bfc\u51fa\u6700\u91cd\u8981\u7684\u4e00\u4e2a pass \u5c31\u662f<code>ToONNX<\/code>\uff0c\u8fd9\u4e2a pass \u4f1a\u8fdb\u884c\u4e00\u4e2a\u6620\u5c04\uff0c\u5c06 TorchScript \u4e2d<code>prim<\/code>\u3001<code>aten<\/code>\u7a7a\u95f4\u4e0b\u7684\u7b97\u5b50\u6620\u5c04\u5230<code>onnx<\/code>\u7a7a\u95f4\u4e0b\u7684\u7b97\u5b50\u3002<\/li><li>\u4f7f\u7528 ONNX \u7684 proto \u683c\u5f0f\u5bf9\u6a21\u578b\u8fdb\u884c\u5e8f\u5217\u5316\uff0c\u5b8c\u6210 ONNX \u7684\u5bfc\u51fa\u3002<\/li><\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u81ea\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/486914187 \u5b98\u7f51\uff1ahttps:\/\/py &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2023\/02\/04\/torchscript\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">TorchScript&#8212;\u6a21\u578b\u90e8\u7f72<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11,26],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12658"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=12658"}],"version-history":[{"count":20,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12658\/revisions"}],"predecessor-version":[{"id":12688,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/12658\/revisions\/12688"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=12658"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=12658"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=12658"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}